Denoising method and device for low-dose CT image

Abstract

The invention relates to a low-dose CT image denoising method and device. The low-dose CT image denoising method comprises the following steps: carrying out image preprocessing on a low-dose CT image; obtaining a first image from the preprocessed low-dose CT image through a convolution sparse coding method, and correspondingly obtaining a first denoising model through training; obtaining a second image from the preprocessed low-dose CT image through an SE attention residual network, and correspondingly obtaining a second denoising model through training; fusing the first image and the second image to generate a final de-noised image, and correspondingly fusing the first de-noising model and the second de-noising model to generate a third de-noising model; and based on the third denoising model, carrying out denoising processing on the low-dose CT image. According to the invention, the denoising processing effect of the low-dose CT image can be effectively improved.

Description

Technical Field

[0001] This invention relates to a method and apparatus for denoising low-dose CT images. Background Technology

[0002] Currently, the main methods for denoising low-dose CT images include projection domain data processing methods and iterative reconstruction methods. Projection domain data processing methods mainly include penalized weighted least squares algorithms, bilateral filtering, and structure-adaptive filtering. However, these methods rely on the original projection data, require deep coupling with the reconstruction algorithm, and have poor versatility, limiting their widespread application in clinical practice. Iterative reconstruction methods are CT image reconstruction methods that gradually approximate the original image by repeatedly optimizing the objective function. However, iterative reconstruction algorithms have high computational complexity, slow reconstruction speed, and are sensitive to parameter settings, making it difficult to balance reconstruction efficiency and image detail. Summary of the Invention

[0003] The purpose of this invention is to provide a method and apparatus for denoising low-dose CT images, which can effectively improve the denoising effect of low-dose CT images.

[0004] Based on the same inventive concept, this invention has two independent technical solutions:

[0005] 1. A method for denoising low-dose CT images, comprising the following steps:

[0006] Step 1: Perform image preprocessing on low-dose CT images;

[0007] Step 2: The preprocessed low-dose CT image is used to obtain the first image through the convolutional sparse coding method, and the first denoising model is obtained accordingly through training; the first image extracts the overall structural features of the low-dose CT image, and the convolutional sparse coding method adopts the alternating iterative multiplier method (ADMM);

[0008] Step 3: The preprocessed low-dose CT image is used to obtain a second image through an SE attention residual network, and a second denoising model is obtained accordingly after training; the high-frequency information of the low-dose CT image is extracted from the second image.

[0009] Step 4: Fuse the first image and the second image to generate the final denoised image, and correspondingly fuse the first denoising model and the second denoising model to generate the third denoising model;

[0010] Step 5: Based on the third denoising model, perform denoising processing on the low-dose CT images.

[0011] Further, in step 2, the first denoising model is represented as follows:

[0012]

[0013] In equation (1), This is a data fidelity item, where S represents a complete low-dose CT image, and d... i Let x represent the i-th filter. i Let represent the local feature map corresponding to the i-th filter, and m represent the total number of filters trained.

[0014] Equation (1) can be written as:

[0015]

[0016] in, D = {D1,D2,L,D} C Let} represent a convolution dictionary with c channels, where each channel represents a filter of size k×k, and λ is a regularization parameter that balances sparsity and fidelity.

[0017] Furthermore, the Alternating Iterative Multiplier Method (ADMM) is used to solve the objective function of equation (2).

[0018] By introducing auxiliary variables into the objective function of equation (2), the problem can be rewritten as an optimization problem in the standard form of the ADMM algorithm:

[0019]

[0020] Equation (3) is solved using the ADMM algorithm, as follows:

[0021]

[0022] u (j+1) =u (j) +x (j+1) -y (j+1) (6)

[0023] Where, x (j+1) y (j+1) u (j+1) represents the update of the convolutional sparse graph, the update of the auxiliary variables, and the update of the Lagrange multipliers in the j-th iteration, respectively.

[0024] Furthermore, in step 2, the convolutional sparse graph x, auxiliary variable y, Lagrange multiplier u, and convolution dictionary D are first initialized.

[0025] During initialization, the preprocessed low-dose CT image is used as the initialization input for the convolutional sparse map x. The initialization of the auxiliary variable y is completed through a module containing two convolutional layers and an activation function. The Lagrange multiplier u is initialized as an all-zero tensor, and the initial values ​​of the dictionary D are composed of predefined convolutional kernel tensors.

[0026] Furthermore, in step 2, before each iteration, the image obtained in the previous iteration is optimized by a convolutional neural network before the current iteration is performed.

[0027] Furthermore, the convolutional neural network is built based on the UNet architecture, with residual modules between adjacent layers.

[0028] Furthermore, during each iteration update, the convolutional sparse map x is updated by first taking the image obtained in the previous iteration and using the Sherman-Morrison formula to provide an initial estimate, which is then used as the input to the convolutional neural network.

[0029] Furthermore, the auxiliary variable y is updated through an activation function; the Lagrange multiplier u is updated by adding the difference between the original variable x and the auxiliary variable y to the current u.

[0030] Furthermore, when training the denoising model, the loss function for each iteration is a weighted combination of L1 loss and perceptual loss, with the ratio of the two values ​​ranging from [0,1].

[0031] 2. A low-dose CT image denoising device for performing the above method.

[0032] The beneficial effects of this invention are as follows:

[0033] This invention obtains a first image from a preprocessed low-dose CT image using a convolutional sparse coding method, and accordingly trains a first denoising model. The preprocessed low-dose CT image is then used to obtain a second image through an SE attention residual network, and accordingly trained a second denoising model. The first and second images are fused to generate the final denoised image, and the first and second denoising models are fused to generate a third denoising model. During the training of the denoising model, before each iteration, the image obtained in the previous iteration is optimized using a convolutional neural network before the current iteration. This invention integrates the feature modeling capabilities of the convolutional sparse coding method (ADMM) and the SE attention residual network, which not only effectively suppresses noise but also better preserves the details and structural information in the image, overcoming the problems of traditional ADMM, such as sensitive parameter selection, high computational cost, and slow convergence speed. In this invention, the SE attention module in the SE attention residual network adaptively adjusts the weight distribution of each channel, thereby strengthening the response of key feature channels. The residual blocks in the network, through skip connections, effectively preserve the structural information of the original image, enhance the expression of useful features, and suppress irrelevant or interfering information, further improving the denoising effect. Compared to most existing deep convolutional neural networks that lack mathematical interpretability, this invention, while preserving the network's expressive power, uses an interpretable optimization iteration process, effectively alleviating the problem of easily lost image details. It successfully completes the denoising task of low-dose CT images with complex noise types and artifacts.

[0034] In each iteration of this invention, the update of the convolutional sparse map x first takes the image obtained in the previous iteration and provides an initial estimate using the Sherman-Morrison formula as the input to the convolutional neural network, providing the network with a more stable and reliable optimization starting point and further ensuring the image processing effect.

[0035] This invention utilizes a convolutional neural network built on the UNet architecture, incorporating residual modules between adjacent layers. The introduction of residual modules between adjacent layers helps the network expand its receptive field while enhancing its feature reuse capability. Each residual module contains an adjustable number of residual units, allowing for flexible configuration based on task complexity. The network employs dense cross-layer connections to further improve feature propagation and reuse efficiency, effectively mitigating the gradient vanishing problem.

[0036] In this invention, the auxiliary variable y is updated through an activation function; the Lagrange multiplier u is updated by adding the difference between the original variable x and the auxiliary variable y to the current u. The update of the auxiliary variable y no longer uses the traditional soft thresholding function, but instead introduces a nonlinear activation function. When the sparse coefficients are non-negative, this substitution does not weaken the signal representation ability, especially when processing positive signals, where the ReLU activation function has an effect similar to the traditional soft thresholding operator. Therefore, using ReLU instead of the soft thresholding function not only maintains the sparsity feature but also enhances the model's expressive power and nonlinear modeling ability. The iterative update of the Lagrange multiplier u follows the update rules in the standard augmented Lagrange multiplier method. The update process of the dictionary D dynamically optimizes each atom in the predefined dictionary during training, enabling it to adaptively learn the local representation features that best fit the target task, thereby improving denoising performance.

[0037] In training the denoising model, this invention uses a weighted combination of L1 loss and perceptual loss as the composite loss function for each iteration, with the ratio of the two values ​​ranging from [0,1]. By fusing these two types of losses, the model can retain more useful details while removing noise, thus jointly training a superior low-dose CT denoising model. When selecting the network layer for calculating the perceptual loss, considering that the feature map processed by the ReLU activation function is very sparse and may limit the model's performance, this invention chooses the feature layer after convolution as the basis for calculating the perceptual loss, rather than selecting the layer after ReLU activation. This more fully preserves feature information, thereby improving training effectiveness and reconstruction quality. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the low-dose CT image denoising method of the present invention;

[0039] Figure 2 This is the overall framework of the network of this invention;

[0040] Figure 3 This is a schematic diagram of the update of the convolutional sparse graph network during each iteration of the present invention;

[0041] Figure 4 This invention is an iterative SE attention residual network;

[0042] Figure 5 This is the SE attention residual network for each iteration of this invention;

[0043] Figure 6 This is a schematic diagram of the residual module in the convolutional sparse graph network of the present invention. Detailed Implementation

[0044] The present invention will now be described in detail with reference to the embodiments shown in the accompanying drawings. However, it should be noted that these embodiments are not intended to limit the present invention. Equivalent changes or substitutions in function, method, or structure made by those skilled in the art based on these embodiments are all within the scope of protection of the present invention.

[0045] Example 1:

[0046] Low-dose CT image denoising methods

[0047] like Figure 1 , Figure 2 As shown, the low-dose CT image denoising method of the present invention includes the following steps:

[0048] Step 1: Perform image preprocessing on low-dose CT images.

[0049] Assuming the size of a low-dose CT image is A×A, several image blocks of size a×a are extracted from it through random cropping as training samples. Combined with data augmentation operations such as image rotation and mirroring, the size of the training dataset is further expanded, thereby improving the robustness and generalization ability of the model.

[0050] Step 2: The preprocessed low-dose CT image is used to obtain the first image through the convolutional sparse coding method, and the first denoising model is obtained accordingly through training; the first image extracts the overall structural features of the low-dose CT image, and the convolutional sparse coding method adopts the alternating iterative multiplier method (ADMM).

[0051] The first denoising model is represented as follows:

[0052]

[0053] In equation (1), This is a data fidelity item, where S represents a complete low-dose CT image, and d... i Let x represent the i-th filter. i Let represent the local feature map corresponding to the i-th filter, and m represent the total number of filters trained.

[0054] Equation (1) can be written as:

[0055]

[0056] in, D = {D1,D2,L,D} C Let} represent a convolution dictionary with c channels, where each channel represents a filter of size k×k, and λ is a regularization parameter that balances sparsity and fidelity.

[0057] The Alternating Multiplier Method (ADMM) is used to solve the objective function of equation (2).

[0058] By introducing auxiliary variables into the objective function of equation (2), the problem can be rewritten as an optimization problem in the standard form of the ADMM algorithm:

[0059]

[0060] Equation (3) is solved using the ADMM algorithm, as follows:

[0061]

[0062] u (j+1) =u (j) +x (j+1) -y (j+1) (6)

[0063] Where, x (j+1) y (j+1) u (j+1) represents the update of the convolutional sparse graph, the update of the auxiliary variables, and the update of the Lagrange multipliers in the j-th iteration, respectively.

[0064] First, initialize the convolutional sparse graph x, auxiliary variable y, Lagrange multipliers u, and convolution dictionary D. During initialization, the preprocessed low-dose CT image is used as the initialization input for the convolutional sparse graph x. The initialization of the auxiliary variable y is completed through a module containing two convolutional layers and an activation function. The Lagrange multipliers u are initialized as all-zero tensors. The initial values ​​of the dictionary D are composed of predefined convolutional kernel tensors. The size of the convolutional dictionary atoms is much smaller than that of the convolutional sparse graph (for example, a 5×5 size convolutional dictionary atom can be selected for a sparse coefficient graph with a resolution of 128×128).

[0065] Before each iteration, the image obtained in the previous iteration is optimized using a convolutional neural network (CNN) before the current iteration. The CNN is built on the UNet architecture and includes residual modules between adjacent layers. During each iteration, the update of the convolutional sparse map x first uses an initial estimate of the image obtained in the previous iteration, obtained through the Sherman-Morrison formula, as input to the CNN. The auxiliary variable y is updated through an activation function. In traditional ADMM methods, the update of the auxiliary variable y typically relies on a soft thresholding function; however, in the deep learning framework, the ReLU function can be used instead of the traditional soft thresholding operation. For non-negative sparse coefficients, this substitution does not affect the signal representation ability; furthermore, when processing positive signals, the ReLU function is approximately equivalent to the element-wise soft thresholding operator. This approach not only effectively updates the auxiliary variable but also enhances the model's expressive power and nonlinear fitting ability. The Lagrange multiplier u is updated by adding the difference between the original variable x (i.e., the convolutional sparse coding graph) and the auxiliary variable y to the current u, thereby gradually approaching the optimal solution.

[0066] Figure 1 , Figure 2 In this context, DADMM is an abbreviation for Deep ADMM, ISER stands for Iterative SE Attention Residual Network, Stage T represents the Tth iteration of DADMM, and Phase S represents the Sth stage of the ISER iteration process.

[0067] In practice, during the t-th iteration, x(t-1) from the previous iteration is used by the Sherman-Morrison formula solver to obtain x'. The result is then fed into a convolutional neural network for further optimization. After selecting a suitable convolutional sparse graph, the auxiliary variable y is updated. y is iteratively updated using an activation function instead of a soft thresholding function. The Lagrange multiplier u follows the standard augmented Lagrange multiplier update rule in each iteration, thus yielding x(t) for that iteration. During training, dictionary D adaptively updates its atoms, gradually learning local structural representation features that conform to the task.

[0068] like Figure 3 , Figure 6As shown, the convolutional neural network adopts a densely connected U-Net structure design, adding residual modules between adjacent layers of the encoder and decoder to enhance feature transfer and information flow efficiency, which helps to extract higher-dimensional and richer image features, improving denoising performance and the model's expressive power. In the U-Net structure, convolutional downsampling and convolutional upsampling operations are used in the downsampling and upsampling stages, respectively, to replace traditional pooling and interpolation methods, thereby further improving the image reconstruction quality. In the input stage of the convolutional neural network, the initial solution is first processed by a residual module containing n sets of k×k convolutional kernels to extract features, generating n feature maps of size a×a. Subsequently, the spatial size of the feature maps is reduced by half through a convolutional downsampling operation and input into a residual module containing 2n sets of k×k convolutional kernels, generating 2n feature maps of size a / 2×a / 2. This process is repeated in each downsampling stage, with the number of feature maps increasing proportionally and the spatial size decreasing step by step, extracting deeper image features. When entering the convolutional upsampling stage of U-Net, the convolutional upsampling operation doubles the spatial size of the feature map in each layer. Simultaneously, in the residual module adjacent to the convolutional upsampling module, the number of channels in the convolutional kernel decreases layer by layer, achieving effective fusion and restoration of multi-scale features. Finally, after the last convolutional upsampling operation, a convolutional sparse map of size a×a and number of channels m is output through a residual module containing m sets of k×k convolutional kernels, thus completing the update of the convolutional sparse map x.

[0069] Where m represents the number of input channels and also the number of atoms in the convolution dictionary; a×a is the image size used during training; n represents the number of convolution kernels; k×k is the size of the convolution kernels; m, n, and k can be flexibly adjusted according to the model convergence (for example, for a 128×128 resolution training image, the number of convolution kernels can be selected in the range of [64, 512], the size of the convolution kernels can be selected in the range of [3, 7], and the number of atoms in the convolution dictionary can be selected to be around 64).

[0070] The weights for convolutional kernels can be initialized randomly, using Gaussian initialization, etc. The batch size of the batch normalization layer can be flexibly adjusted based on model convergence to obtain better training results. The activation function can be one of the following commonly used functions: sigmoid, ReLU, Leaky ReLU, or softplus; the specific choice can be adjusted based on the network structure.

[0071] Step 3: The preprocessed low-dose CT image is used to obtain a second image through the SE attention residual network, and a second denoising model is obtained accordingly after training; the high-frequency information of the low-dose CT image is extracted from the second image.

[0072] like Figure 4 , Figure 5 As shown, the Iterative SE Attention Residual Network is a feature enhancement module that integrates an iterative strategy, an SE channel attention mechanism, and a residual learning structure. First, the input low-dose CT image is processed through preliminary convolutional operations to extract a high-dimensional feature space, providing a rich representational foundation for subsequent deep networks. Then, in each iteration, the feature map is sequentially passed through two residual blocks embedded with the SE channel attention mechanism to achieve effective fusion of local and global features. During this process, the SE attention module adaptively learns the importance weights between channels, guiding the network to focus more on key channels with structural and textural information. Simultaneously, to enhance the network's nonlinear expressive power, the model uses the ReLU activation function to perform element-wise nonlinear suppression on intermediate features, effectively reducing the interference of noise responses. The updated feature map in each iteration is fused with the initial features through a residual approach, enabling the network to extract useful information more accurately in each iteration and gradually refine the image texture over multiple iterations.

[0073] Step 4: Fuse the first image and the second image to generate the final denoised image, and accordingly fuse the first denoising model and the second denoising model to generate the third denoising model.

[0074] The overall structural features of the image restored by deep ADMM and the high-frequency detail information extracted by the iterative SE attention residual mechanism are effectively fused to balance the global consistency and local details of the image.

[0075] In each training iteration, the model uses a weighted combination of L1 loss and perceptual loss as a composite loss function, with the weight ratio of the two ranging from [0, 1]. L1 loss, by calculating the mean absolute error between the predicted and real images, effectively suppresses noise and improves the overall denoising result. Perceptual loss, on the other hand, measures the differences between images at the feature level, helping to capture structural details and texture information, thus enhancing the perceptual quality of the image. By fusing these two types of losses, the model can retain more useful details while removing noise, thereby jointly training a superior low-dose CT denoising model. When selecting the network layer for calculating the perceptual loss, considering that the feature map after ReLU activation is very sparse and may limit the model's performance, this invention chooses the convolutional feature layer as the basis for calculating the perceptual loss, rather than selecting a layer after ReLU activation, thus more fully preserving feature information and improving training effectiveness and reconstruction quality.

[0076] Network parameters can be updated using optimization algorithms such as Stochastic Gradient Descent (SGD) and Adam. During training, one or more of the hyperparameters, such as the learning rate, number and size of convolutional kernels, weight parameters, and number of network layers, can be dynamically adjusted based on the network's convergence and test results to optimize model performance.

[0077] Step 5: Based on the third denoising model, perform denoising processing on the low-dose CT images.

[0078] After training the model, 35 untrained low-dose CT images were selected as a test set for testing, generating corresponding denoised images. Subsequently, for each denoised image, subjective visual effect analysis and objective quantitative evaluation were performed. For quantitative analysis, commonly used image quality metrics such as PSNR and SSIM were used to evaluate the denoising effect, comprehensively measuring the model's ability to restore image structure and preserve details. Finally, the evaluation results of all test images were averaged to comprehensively evaluate the model's overall performance and practical application effectiveness in low-dose CT denoising tasks.

[0079] Example 2:

[0080] Low-dose CT image denoising device

[0081] The low-dose CT image denoising device is used to perform the low-dose CT image denoising method described above.

[0082] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

[0083] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

Claims

1. A method for denoising low-dose CT images, characterized in that, Includes the following steps: Step 1: Perform image preprocessing on low-dose CT images; Step 2: The preprocessed low-dose CT image is used to obtain the first image through a convolutional sparse coding method, and the first denoising model is obtained accordingly through training; the convolutional sparse coding method adopts the alternating iterative multiplier method (ADMM). Step 3: The preprocessed low-dose CT image is used to obtain a second image through the SE attention residual network, and a second denoising model is obtained accordingly through training. Step 4: Fuse the first image and the second image to generate the final denoised image, and correspondingly fuse the first denoising model and the second denoising model to generate the third denoising model; Step 5: Based on the third denoising model, perform denoising processing on the low-dose CT images.

2. The method for denoising low-dose CT images according to claim 1, characterized in that, In step 2, the first denoising model is represented as follows: In equation (1), This is a data fidelity item, where S represents a complete low-dose CT image, and d... i Let x represent the i-th filter. i Let represent the local feature map corresponding to the i-th filter, and m represent the total number of filters trained. Equation (1) can be written as: in, D = {D1,D2,L,D} C Let} represent a convolution dictionary with c channels, where each channel represents a filter of size k×k, and λ is a regularization parameter that balances sparsity and fidelity.

3. The low-dose CT image denoising method based on convolutional sparse coding according to claim 2, characterized in that: The Alternating Multiplier Method (ADMM) is used to solve the objective function of equation (2). By introducing auxiliary variables into the objective function of equation (2), the problem can be rewritten as an optimization problem in the standard form of the ADMM algorithm: Equation (3) is solved using the ADMM algorithm, as follows: u (j+1) =u (j) +x (j+1) -y (j+1) . (6) Where, x (j+1) y (j+1) u (j+1) denoted as the update of the convolutional sparse graph, the update of the auxiliary variables, and the update of the Lagrange multipliers in the j-th iteration, respectively, with ρ being the regularization parameter.

4. The method for denoising low-dose CT images according to claim 3, characterized in that: In step 2, the convolutional sparse graph x, auxiliary variable y, Lagrange multiplier u, and convolution dictionary D are first initialized. During initialization, the preprocessed low-dose CT image is used as the initialization input for the convolutional sparse map x. The initialization of the auxiliary variable y is completed through a module containing two convolutional layers and an activation function. The Lagrange multiplier u is initialized as an all-zero tensor, and the initial values ​​of the dictionary D are composed of predefined convolutional kernel tensors.

5. The method for denoising low-dose CT images according to any one of claims 1 to 4, characterized in that: In step 2, before each iteration, the image obtained in the previous iteration is optimized by a convolutional neural network before the current iteration is performed.

6. The method for denoising low-dose CT images according to claim 5, characterized in that: The convolutional neural network is built on the UNet architecture, with residual modules between adjacent layers.

7. The method for denoising low-dose CT images according to claim 5, characterized in that: During each iteration, the convolutional sparse map x is updated by first taking the image obtained in the previous iteration and using the Sherman-Morrison formula to provide an initial estimate, which is then used as the input to the convolutional neural network.

8. The method for denoising low-dose CT images according to claim 3, characterized in that: The auxiliary variable y is updated through an activation function; the Lagrange multiplier u is updated by adding the difference between the original variable x and the auxiliary variable y to the current u.

9. The method for denoising low-dose CT images according to claim 1, characterized in that: When training the denoising model, the loss function for each iteration is a weighted combination of L1 loss and perceptual loss as the composite loss function, and the ratio of the two values ​​ranges from [0,1].

10. A low-dose CT image denoising device, characterized in that, Used to perform the method according to any one of claims 1-9.

Images